Module outline Agent Communicationpift6802/pdf/03_ACL.pdf1 Agent Communication Kostas Stathis...

1

Agent Communication

Kostas Stathis

[email protected]

www.soi.city.ac.uk/~kostas

Kostas Stathis © 2002 + Laurent Magnin © 20052

Module outline

ζ Motivation: What is an agent

ζ Agent Infrastructureψ Architectures

ψ Mobile Agents

ζ Components of an agentψ Planning

ψ Learning

ζ Coordination of agentsψ Agent communication

ψ Negotiation

ζ Applicationsψ Information agentsψ Guest lecture


Learning outcomes

ζ Appreciate why communication is important inMAS;

ζ Learn the basics of Speech Act Theory;ζ Understand what is an Agent Communication

Language (ACL);ζ Compare and contrast KQML and FIPA, with focus

on semantics;ζ Become aware of agent communication

technologies.


Overview

ζ Motivation

ζ Communication

ζ Speech Act Theory

ζ Agent Communication Languages (KQML,FIPA, Interaction Protocols)

ζ Comparison of KQML and FIPA

ζ Summary


Motivation

(Social) Agents must be able to communicate

in order to support:

ζ user-agent communication and interaction;

ζ co-operative problem solving;

ζ action co-ordination;

ζ inter-operation.


Communication

The definition below is due to Russel & Norvig

(1995) book on AI:

“ In general, communication is the intentionalexchange of information brought about by theproduction and perception of signs drawn from ashared system of conventional signs.”

2


Human communication

Humans:

ζ use a limited number of conventional sign(smiling, shaking hands, etc);

ζ have developed a complex, structuredsystem of signs known as language.


Types of meaning

The formal study of language provides for:ψ how will the communicative actions be structured

(syntax);

ψ what will the symbols of communicative actions denote(semantics);

ψ how will the communicative action symbols beinterpreted (pragmatics).

Can we analyse human communication in

order to build communicating agents?


Communication as action

Communication is often interpreted aslanguage action (or communicative act):

ζ acts are described by episodes ;ζ each episode assumes a speaker S wanting to

achieve goal G and, as a result, conveysproposition P to hearer H using communicative actA.

Note: we take here an intentional stance.


Communication as action (cntd)

The speaker engages in the activities of:

ζ Intention - S wants that H believes (or does) P,because S believes (or wants) that P is (or to be)the case.

ζ Generation - S chooses the signs denoting P.

ζ Synthesis - S utters A having P as content.

E.g. John, please close the door.


Communication as action (cntd)

The activities at the hearer H include:

ζ Perception - H perceives A’ (ideally A = A’, butmisconception is possible).

ζ Analysis - H infers that A’ has possible meaningsP1..Pn (propositions can have several meanings).

ζ Disambiguation - H infers that S intended to conveyPi (ideally Pi = P, but misinterpretation is possible).


Communication as action(cntd)

Processes at the hearer H include (cntd):

ζ Incorporation - H decides to believe Pi (or rejects itif out of line with what H believes).

Can we formally generate, analyse, anddisambiguate communicative acts?

3


Speech act theory

Based on the work by Austin*:

Speech act theories focus on how language isused by people in their every day activities, toachieve their goals/intentions (also known aspragmatic theories).

*Austin (1962), How to Do Things with Words.


Speech act theory

Consider the act “Shut the door!”. We call:

ζ the locution - the physical utterance with context andreference, i.e. who is the speaker and the hearer, whichdoor, etc;

ζ the illocution - the act of conveying intentions, i.e., speakerwants the hearer to close the door;

ζ the perlocution - the action that occur as a result of theillocution, i.e. hearer closes the door.


VF…

ζ (http://www.episteme.u-bordeaux.fr/Dico1/actes.htm)

ζ Par la suite, Austin refonde sa typologie en troiscatégories d'actes de langage :ψ 1. Les actes locutionnaires qui correspondent au fait

de dire quelque chose.ψ 2. Les actes illocutionnaires accomplis en disant

quelque chose.ψ 3. Les actes perlocutionnaires accomplis par le fait

de dire quelque chose.


Speech act classification

Searle* identified different types of

speech acts:

ζ representatives - tell others how things are, e.g.“John is English”;

ζ directives - try to get others do things, e.g. “Shutthe door!”;

*Searle (1969), Speech Acts, Cambridge University Press.


Speech act classification(cntd)

ζ commisives - commit ourselves to future actions,e.g. “I promise to marry you”;

ζ expressives - express our feelings and attitudes,e.g. “I am happy”;

ζ declarations - bring about changes through ourutterances, e.g. “I pronounce you man and wife”.

Debatable which classification is appropriate.


Speech act components

In general a speech act can be seen to

have two main components:

ζ a performative verbe.g. inquire, inform, …;

ζ the propositional content e.g. “shut the door”.

4


Speech act examples

ζ Informing that the door is shut:ψ performative=inform

ψ content= “(shut door)” or “shut(door)”

ψ speech act= “The door is shut!”

ζ Inquiring if the door is shut:ψ performative=inquire

ψ Content= “(shut door)”

ψ speech act= “Is the door shut?”

ζ Orderψ Performative = “request”, content= “(shut door)”


Agent Communication

Speech acts form messages in an Agent

Communication Language (ACL).

ζ An ACL is a high-level language whose primitives andstructures are expressly tailored to support the exchange ofmessages amongst multiple artificial agents;

ζ An ACL exists in a logical layer on top of existinginfrastructures such as TCP/IP, HTTP, or IIOP.


Agent Communication (cntd)

The two prevalent ACLs are:

ζ KQML (Knowledge Query and Manipulation Language);

ζ FIPA ACL (Foundation of Intelligent Physical Agents).


KQML

Developed by the (D)ARPA knowledge sharing initiative.

ζ based on the idea that communication requires a commonlanguage;

ζ requires that common language is divided into syntax,semantics and pragmatics, using:

ψ KIF - syntax of content;

ψ Ontoligua - semantics of content;

ψ speech acts - pragmatics of communication;


KQML: Syntax

KIF (Knowledge Interchange Format) :

ζ is based on first-order logic set theory;

ζ encodes declarative knowledge;

ζ Lisp like syntax, e.g.:ψ (forall? x(=>(P ?x)(Q ?x)))

ψ (=>(apple ?x)(red ?x))

ψ (= (size chip1) 20)

ζ see http://logic.stanford.edu/kif/dpans.html for moredetails.


KQML: Semantics

Ontology: a common vocabulary and agreed upon

meanings to describe a subject domain.

ζ Ontolingua is a language for building, publishing, andsharing ontologiesψ ontologies can be automatically translated into other content

languages, including KIF, Prolog, etc.

ψ the main language included primitives for combining ontologies.

ζ see www.ksl.stanford.edu/software/ontolingua/ for moredetails.

ζ Also protégé.stanford.edu : Ontology editor

5


KQML: Semantics (cntd)

fruit

apple lemon orange

fruit

apple citrus pear

lime lemon orange

Example (by Finin and Lambrou)


KQML: Pragmatics

ζ The pragmatics define the performative verbs, forexample:ψ ask-if (‘is it true that…’);

ψ perform (‘please perform the following action…);

ψ tell (‘it is true that …’);

ψ reply (‘the answer is …’);

More on pragmatics later (see ACL semantics p38).


A KQML Message

Msg above represents a single KQML speech actdescribed by a list of attribute/value pairs e.g. :content,:language, :from, :in-reply-to.

(tell :sender bookShopAgent123 :receiver ksAgent :in-reply-to id7.34.96.45391 :ontology books :language Prolog :content “price(ISBN3429459,24.95)”)

performative

parameter

value


A KQML Dialogue

Consider agents A and B “talking” about the prices

of books bk1 and bk2:

A to B: (ask-if (> (price bk1) (price bk2)))B to A: (reply true)B to A: (inform (= (price bk1) 25.50))B to A: (inform (= (price bk2) 19.99))

For convenience msg format above is simplified andattribute/value pairs for :ontology etc. are omitted.


FIPA

FIPA (Foundations of Intelligent Physical Agents)

ζ its purpose is to promote the success of emergingagent-based applications, services and devices.

ζ its goal is pursued by making available in a timelymanner, internationally agreed specifications forinteroperable agent-based applications, services anddevices.

ζ see http://www.fipa.org/ for more details.


Agent Management in FIPA

FIPA provides a reference

model for agent:

ζ creation;

ζ registration;

ζ location;

ζ communication;

ζ migration; and

ζ retirement.

6


FIPA ACL

ζ Like KQML, FIPA ACL is basedon idea of Speech Acts (seeperformatives below).

ζ agree

ζ cancel

ζ cfp

ζ confirm

ζ disconfirm

ζ failure

ζ inform

ζ inform-if

ζ inform-ref

ζ not-understood

ζ propose

ζ query-if

ζ query-ref

ζ refuse

ζ reject-proposal

ζ request

ζ request-when

ζ request-whenever

ζ subscribe


Ontologies in FIPA

OKBC

Agent

OntologyServer

OQL

OntologyServer

OntologyAgent

(Ontolingua) (DB of ODL definitions)

http

OntologyServer

( XML )

FIPAcomponents

DFAgent

Ontology designer

OA-2

Non-FIPAcomponents

ACL Channel

OntologyAgentsprovide ontologyservices, such astranslation, to FIPA agents

Ontologyserversprovideshared

ontologiesto agents andother systems


FIPA ACL Example

(inform :sender agent1 :receiver hpl-auction-server (price (bid good02) 150) :in-reply-to round-4 bid04 :language sl :ontology hpl-auction)

ACL message

:content

:reply-with

Note: FIPA uses adifferent content language from KQML called SL (seelater), which is based on modal logic.


FIPA Interaction Protocols

Ongoing conversation between agents often fall in

typical patterns:

ζ typical patterns of message exchange are calledinteraction protocols ;

ζ FIPA pre-specifies a number of protocols.

Advantage: ease of implementation (an agentcan engage in meaningful interactions simplyby following an interaction protocol).


Example: Query-Protocol

The protocol states that:ζ IF an agent asks another agent

if a proposition holds or not

ζ THEN the participant mustreply with answer informing theinitiator, or may refuse toanswer, or fail to answer, orfailing to understand thequestion.


FIPA Transport management

ζ a platform provides to each agent achannel;

ζ a transfer protocol is used to transfermessages between channels;

ζ the transport service deliversmessages within a platform orbetween platforms;

ζ ACL is the payload of the transportservice and protocol.

ζ the envelope contains transportinformation.

7


Using FIPA ACL (Java-based)

ACLMessage incoming;ServiceAgent me;...%suppose that agent me has received an incoming request about whether% the price of bk1 is £25, and wants to reply that this is true. Here is an% example of how to formulate reply using a FIPA like platform:

ACLMessage reply = new ACLMessage(”inform");reply.setDest(msg.getSource());reply.setSource(me.getName());reply.setContent("true");reply.setReplyTo(msg.getReplyWith());reply.setProtocol("fipa-request");reply.setOntology(Constants.ONTOLOGY);reply.setLanguage(Constants.LANGUAGE);me.send(reply);...


Comparing ACLs

In developing agent communication one needs

to know: Which ACL is better?ψ FIPA?

ψ KQML?

ψ Some other?

Comparison involves, amongst other things, asking

how meaning is assigned to an ACL’s speech acts.


ACL Semantics

We will examine the semantics of:

ψ KQML;

ψ FIPA;

ψ KQML vs FIPA;

Semantics: Not about what does the content of

an act denotes, but how should the act be

interpreted by an agent.

KQML Semantics: Objectives

ζ to capture intuitions expressed about KQML;

ζ to avoid commitments to specific system designs andarchitectures, agent theories, problem-solving strategies,etc.

ζ to offer a formalism that might be understood by peoplethat are not logicians;

ζ to (loosely) model performatives as speech acts.

Kostas Stathis © 200240

KQML semantics: Assumption

Describing the state of an agent before sending a

message and after receiving it is a useful basis for

ascribing meaning to the communicative acts

(Lambrou 1996, PhD Thesis). ⇓

A two-step process:

ψ Which agents’ states?

ψ Having a language to describe agents’ states.

Kostas Stathis © 200241 Kostas Stathis © 2002 + Laurent Magnin © 200542

KQML semantics: which agentstates?

ζ Preconditions - indicate the necessary state for an agentin order to send a performative and for the receiver toaccept it and successfully process it.

ζ Postconditions - describe the states of both participantsafter the successful utterance of a performative (by thesender) and after the receipt and processing (but beforea counter utterance) of a message (by the receiver).

8

KQML Example: TELL(A,B,P)

A tells B that A believes P to be true (for A).

ζ Pre(A): bel(A,X) ∧ know(A, want(B, know(B,S)))and S = bel(B,X) or ¬ bel(B,X)

Pre(B): intend(B, know(B, S))ζ Post(A): know(A, know(B, bel(A,X))) Post(B): know(B, bel(A,X))ζ Completion: know(B, bel(A,X))ζ The completion and post conditions hold unless a SORRY

or ERROR suggests B’s inability to properly acknowledgethe TELL.

Kostas Stathis © 200243 Kostas Stathis © 2002 + Laurent Magnin © 200544

FIPA ACL Semantics

Define semantics of performatives in terms of:

ζ Feasibility Preconditions - define the conditions that ought tobe true before an agent may plan to execute acommunicative act;

ζ Rational Effects - the effects that an agent hopes to bringabout by performing an action (but with no guarantee thatthey will be achieved).

FPs and REs involve agents state descriptions in SL.


Semantic Language (SL)

ζ SL is the formal language used to define the semantics ofFIPA ACL;ψ logical propositions are expressed in a logic of mental attitudes

and actions;

ψ the logical framework is a first order modal language with identity(similar to Cohen & Levesque);

ψ SL provides formalizations for three primitive mental attitudes:Belief, Uncertainty and Choice (or Goal);

ζ SL can express propositions, objects and actions.


BUC(http://www.fipa.org/specs/fipa00008/SC00008I.html)

ζ (B <agent> <expression>)Belief. It is true thatagent believes that expression is true.

ζ (U <agent> <expression>)Uncertainty. It is truethat agent is uncertain of the truth of expression.Agent neither believes expression nor its negation,but believes that expression is more likely to be truethan its negation.

ζ (I <agent> <expression>)Intention. It is true thatagent intends that expression becomes true and willplan to bring it about.


FIPA ACL Semantics: Inform

i informs j about φ being true.

ζ <i, inform( j, φ )>FP: Biφ ∧ ¬ Bi( Bifjφ ∨ Uifjφ)

[ FP: Biφ ∧ ¬ Bi(Bjφ ∨ Ujφ) ]

RE: Bjφ

ζ The sender:• believes that proposition φ is true;• intends that receiver j also comes to believe that φ is true;• does not already believe receiver has any knowledge

about φ .

ζ It’s up to receiver to adopt the belief.


KQML vs FIPA: differences?

ζ Different semantics; mapping performatives between FIPAand KQML is a fruitless exercise.

ζ Different treatment of the “administration primitives”; in FIPAACL register, unregister, etc., are treated as requests withreserved meaning;

ζ FIPA ACL has no facilitation primitives, e.g., broker,recommend, recruit, etc.;

ζ KQML does not commit to a content language while FIPAdoes.

9


KQML vs FIPA: which one?

Finin and Lambrou argue that:

ζ programmers do not care about semantics, especiallywhen applications do not require modalities (belief,intention, etc.);

ζ similar syntax helps that a developer will not have toalter the code about messages.

ζ the code that processes the primitives should changedepending on whether the code follows the propersemantics.


KQML vs FIPA: which one?(cntd)

ζ FIPA ACL is more powerful with composing newprimitives (power due to SL language as a contentlanguage to describe agents’ states).

ζ KQML’s weakness is its non-commitment to a contentlanguage.

ζ Both have shortcomings; see references at the end ofthe lecture.


Alternative approaches

ζ CORBA;

ζ Java (RMI, EJB, Jini);

ζ OLE/COM/DCOM/ActiveX (Microsoft);

ζ SOAP (XML-based);

ζ Service-based platfroms (e.g. JXTA);

ζ Services-based languages (e.g. e-Speak).


Summary

ζ Communication

ζ Speech act theory

ζ Agent communication languagesψ KQML

ψ FIPA

ζ Comparison of KQML and FIPA

ζ Alternative approaches


Websites

KQML http://www.cs.umbc.edu/kqml/

KIF http://logic.stanford.edu/kif/

Ontolingua http://www-ksl-svc.stanford.edu:5915/&service=frame-editor

FIPA http://www.fipa.org/

DAML http://www.daml.org/


Further reading

[1] M. Wooldrige. Reasoning about Rational Agents. The MIT Press, 2000,Chapter 7

[2] M. Huhns, L. Stephens. Multiagent systems and societies of agents. InMultiagent Systems - A Modern Approach to Distributed ArtficialIntelligence, G. Weiss (Ed.), The MIT Press, 2001, p.79-120.

[3] F. Dignum and M. Greaves. Issues in Agent Communication: AnIntroduction. In F. Dignum and M. Greaves(eds.) Issues in AgentCommunication (LNCS-1916), Springer-Verlag, 2000,pages 1-16.http://www.cs.uu.nl/~dignum/papers/papers.html

[4] Y. Labrou, T. Finin and Y. Peng, The current landscape ofAgentCommunication Languages, Intelligent Systems, volume 14, number2, March/April 1999, IEEE Computer Society. http: umbc.edu/~finin/papers/

[5] M. P. Singh. Agent Communication Languages: Rethinking the Principles.http://www.csc.ncsu.edu/faculty/mpsingh/papers/index.html

10


Further reading (cntd)

[6] Jeremy Pitt, Abe Mamdani: Some Remarks on the Semantics of FIPA?sAgent Communication Language. Autonomous Agents and Multi-AgentSystems 2(4): 333-356 (1999)

[7] Jeremy Pitt, Abe Mamdani: Designing Agent Communication Languages forMulti-agent Systems. MAAMAW 1999: 102-114

[8] Cohen, P. R. and Levesque, H. J. (1990b). Rational interaction as the basisfor communication. In Cohen, P. R., Morgan, J., and Pollack, M. E., editors,Intentions in Communication, pages 221-256. The MIT Press: Cambridge,MA.

Module outline Agent Communicationpift6802/pdf/03_ACL.pdf1 Agent Communication Kostas Stathis...

Documents

Transcript of Module outline Agent Communicationpift6802/pdf/03_ACL.pdf1 Agent Communication Kostas Stathis...